nx_langcensus.nx

nx_langcensus.nx -- THE WHOLE-POPULATION COMPILER CENSUS: compile EVERY top-level program in the tree with a GIVEN compiler, classify each outcome by REASON, and (against a baseline census) name every program whose class changed. This is the instrument a language or codegen change is measured by. WHY THIS EXISTS, MEASURED 2026-09-03. The estate's incumbent compiler ruler, nx_cc_equiv_gate, runs EQ_ROWS = 10 corpus rows. Ten rows is a SAMPLE, and a sample can see that a mechanism works; it can never see that a CLASS of program stopped compiling. On the day LN36/LN39/LN40 landed, a first cut of the integer literal bound refused every decimal above i64 max. Every hand-written fixture was GREEN and the 10-row equivalence gate was GREEN, while 286 real estate programs had stopped compiling -- the estate writes u64 BIT PATTERNS in decimal (FNV offset bases, all-ones masks, hash multipliers), so the correct bound is 2^64-1. Only a census over all 13,790 programs could see that class, and it existed only as a shell loop. ★★★★★★A RULER THAT SAMPLES CAN PROVE A MECHANISM AND CAN NEVER PROVE A POPULATION -- AND A COMPILER CHANGE IS A CLAIM ABOUT THE POPULATION. THREE OUTCOMES, NEVER TWO. A compiler that LOOPS on an input is not a compiler that refused it: before the LN40 packer fix a 20-digit fraction literal spun forever, and a census without a per-program deadline hangs at that file and reports nothing about the 9,000 programs after it. HANG is its own class here, and it is counted, named, and treated as a regression. A harness failure (fork/pipe/exec) is a FOURTH class that asserts nothing about the subject -- ★AN AXIS THAT CANNOT SEE MUST ABSTAIN, NEVER ACQUIT. THE REASON TRAVELS WITH THE COUNT. Every refusal row carries the compiler's own first diagnostic line with the file:line prefix stripped, so rows group into causes and a 773-refusal census reads as a dozen classes rather than a number. The delta against a baseline names each program, because ★A COUNT WITHOUT A WORKLIST IS NOT ACTIONABLE. NO SILENT CAP. If the file table or any buffer fills, coverage_complete goes to 0 and the organ REFUSES (exit LC_EXIT_UNPROVEN) rather than publishing a prefix as a population. Usage: nx_langcensus <cc.elf> <out-path> [timeout_ms] [baseline-census-path] Run with CWD = nxc2/ (or any root holding runtime/). Writes one row per program to <out-path>: OK <path> | ERR <path> :: <reason> | HANG <path> | HARN <path> rc=<n> Exit: 0 NO-REGRESSION - 1 REGRESSION (new refusals or hangs, each named) - 2 usage - 3 UNPROVEN. With no baseline the exit is 0 whenever the census completed: a first census has nothing to regress from, and saying otherwise would make the instrument's first run a permanent RED. RESOURCE ENVELOPE. One compiler fork per program, strictly sequential, each bounded by timeout_ms. Buffers (capture, dirent, path table, reason table) are allocated ONCE before the walk and reused -- ★NEVER ALLOCATE IN A HOT LOOP: a per-row capture buffer over 13,790 rows is gigabytes of address space for nothing. The compiler's stdout is discarded to the capture buffer and never to disk.